MobilEye announces plans for robotaxis in Munich


Companies are at the stage of announcing real pilot projects for robotaxi service. Now MobilEye announces they will start a robotaxi service in Munich and Tel Aviv by 2022. What are the new metrics of success for a team?

See more at MobilEye announces plans for robotaxis in Munich


deepai dot org/publication/safe-perception-a-hierarchical-monitor-approach

Perception systems of automated vehicles rely on AI-based detection algorithms. In general, these algorithms offer great detection quality, nevertheless they rely on the quality of the training data and thus there is always the possibility that objects are missed even though they are clearly visible, simply as these were not covered in any training data.

Consequently, it is hard to construct a proof that such an AI-based system is safe under all operating conditions. To address that gap, we presented a novel monitor to check the correctness of a given perception system. The monitor uses LiDAR information, and is realized as a combination of a rule-based probability filter, that allows the system to correctly identify prominent objects that may have been missed by the primary perception system, and additional filters that both increase detection rate and decrease false alarm rate.

We highlighted in our evaluation that a state-of-art AI detector for LiDAR perception can miss prominent and safety relevant objects in the surrounding of the vehicle. Also these detectors are sensitive to configuration parameters. A small change in the configuration of the confidence threshold can lead to miss detection of clearly visible objects. The use of lower confidence can reduce the miss detections, but only at the cost of more false alarms.

With our monitor architecture we are able to eliminate the majority of detection misses, and the few remaining ones are all heavy occluded and not in direct connection to the ego vehicle (thus not safety relevant). In addition to that the false alarm rate is still low ( <2%).

Hence, in summary, with our approach we can improve safety of the LiDAR perception system and still maintain a high availability. It is worth mentioning, that in this work we focused on a single channel LiDAR system. In a final AV, it is desirable to use additional modalities in parallel, to gain robustness against sensor failures.
Intel Labs Europe

Add new comment